A semanticaly versioned Resque plugin which ensures for a given queue, that only one worker is working on a job at any given time.
Resque::LonelyJob differs from resque-queue-lock, resque-lock and resque-loner in that the same job may be queued multiple times but you're guaranteed that first job queued will run to completion before subsequent jobs are run.
However, it is a very strong possibility that subsequent jobs are re-ordered due to the implementation of reenqueue. (See Example #2 for an alternative approach that attempts to preserve job ordering but introduces the possibility of starvation.)
Therefore it is recommended that the payload for jobs be stored in a separate redis list distinct from the Resque queue (see Example #3).
Version 1.x Requires Resque >= 1.20.0 and < 1.25.0.
Requires a version of MRI Ruby >= 1.9.3.
Add this line to your application's Gemfile:
gem 'resque-lonely_job', '~> 1.0.0'
And then execute:
$ bundle
Or install it yourself as:
$ gem install resque-lonely_job
require 'resque-lonely_job'
class StrictlySerialJob
extend Resque::Plugins::LonelyJob
@queue = :serial_work
def self.perform
# only one at a time in this block, no parallelism allowed for this
# particular queue
end
end
Let's say you want the serial constraint to apply at a more granular level. Instead of applying at the queue level, you can overwrite the .redis_key method.
require 'resque-lonely_job'
class StrictlySerialJob
extend Resque::Plugins::LonelyJob
@queue = :serial_work
# Returns a string that will be used as the redis key
# NOTE: it is recommended to prefix your string with the 'lonely_job:' to
# namespace your key!
def self.redis_key(account_id, *args)
"lonely_job:strictly_serial_job:#{account_id}"
end
# Overwrite reenqueue to lpush instead of default rpush. This attempts to
# preserve job ordering but job order is *NOT* guaranteed and also not
# likely. See the comment on SHA: e9912fb2 for why.
def self.reenqueue(*args)
Resque.redis.lpush("queue:#{Resque.queue_from_class(self)}", Resque.encode(class: self, args: args))
end
def self.perform(account_id, *args)
# only one at a time in this block, no parallelism allowed for this
# particular redis_key
end
end
NOTE: Without careful consideration of your problem domain, worker starvation and/or unfairness is possible for jobs in this example. Imagine a scenario where you have three jobs in the queue with two resque workers:
+---------------------------------------------------+
| :serial_work |
|---------------------------------------------------|
| | | | |
| redis_key: | redis_key: | redis_key: | ... |
| A | A | B | |
| | | | |
| job 1 | job 2 | job 3 | |
+---------------------------------------------------+
^
|
Possible starvation +-----------+
for this job and
subsequent ones
When the first worker grabs job 1, it'll acquire the mutex for processing redis_key A. The second worker tries to grab the next job off the queue but is unable to acquire the mutex for redis_key A so it places job 2 back at the head of the :serial_work queue. Until worker 1 completes job 1 and releases the mutex for redis_key A, no work will be done in this queue.
This issue may be avoided by employing dynamic queues, https://blog.kabisa.nl/2010/03/16/dynamic-queue-assignment-for-resque-jobs/, where the queue is a one to one mapping to the redis_key.
The secret to preserving job order semantics is to remove critical data from the resque job and store data in a separate redis list. Part of a running job's responsibility will be to grab data off of the separate redis list needed for it to complete its job.
+---------------------------------------------------+
| :serial_work for jobs associated with key A |
|---------------------------------------------------|
| data x | data y | data z | ... |
+---------------------------------------------------+
+---------------------------------------------------+
| :serial_work for jobs associated with key B |
|---------------------------------------------------|
| data m | data n | data o | ... |
+---------------------------------------------------+
+---------------------------------------------------+
| :serial_work |
|---------------------------------------------------|
| | | | |
| redis_key: | redis_key: | redis_key: | ... |
| A | A | B | |
| | | | |
| job 1 | job 2 | job 3 | |
+---------------------------------------------------+
It now doesn't matter whether job 1 and job 2 are re-ordered as whichever goes first will perform an atomic pop on the redis list that contains the data needed for its job (data x, data y, data z).
The behavior when multiple jobs exist in a queue protected by resque-lonely_job is for one job to be worked, while the other is continuously dequeued and requeued until the first job is finished. This can result in that worker process pegging a CPU/core on a worker server. To guard against this, the default behavior is to sleep for 1 second before the requeue, which will allow the cpu to perform other work.
This can be customized using a @requeue_interval class instance variable
in your job like so:
require 'resque-lonely_job'
class StrictlySerialJob
extend Resque::Plugins::LonelyJob
@queue = :serial_work
@requeue_interval = 5 # sleep for 5 seconds before requeueing
def self.perform
# some implementation
end
end
- Fork it
- Create your feature branch (
git checkout -b my-new-feature) - Commit your changes (
git commit -am 'Added some feature') - Push to the branch (
git push origin my-new-feature) - Create new Pull Request